Copy

cp:

$ cp -a fromdir todir

-a, --archive: same as -dR --preserve=all.

cpio:

$ cd fromdir
$ find . | cpio -pdumv todir

tar:

$ cd fromdir; tar cf - . | (cd todir; tar xfp -)

tar with compression and ssh tunnel:

$ tar cvf - fromdir | gzip -9c | ssh user@host 'cd todir; gzip -cd | tar xpf -'

dd and aiocp performance compare:

# time dd if=/dev/zero of=/dev/sda3 bs=1M count=30k oflag=direct
30720+0 records in
30720+0 records out
32212254720 bytes (32 GB) copied, 167.305 seconds, 193 MB/s

real    2m47.320s
user    0m0.019s
sys     0m4.176s

# time ./aiocp -s 30G -w -b 1M -n 4 -f DIRECT /dev/sda3

real 1m59.322s
user 0m0.013s
sys 0m3.337s

Why so much faster? Because dd reads 1MB, writes 1MB, waits, reads 1MB, writes 1MB, waits, etc. Even though the read from /dev/zero is basically just a memset in the kernel, it introduces enough latency in the process that the write stream stalls.